Browse Courses
|
Gen AI Ready
|
Prologue
|
Scene 6 – Tool Training Lab: Machine Learning (learning from examples)

Scene 6 - Tool Training Lab: Machine Learning (learning from examples)

by Carmen Smith

Scenario: The Pattern Engine Trial

Operator Briefing: Welcome to modern AI training. You won’t write rules, you’ll train a model from examples. Under the hood, this is a small neural network learning patterns from your data. Your job is to prove the most important truth in machine learning: a model can look smart when the world stays the same… then fail the moment conditions change. Then you’ll “level it up” by improving the training data (this is what scale really means at your level).

Your mission: Train your Close vs Far model, test it, change one condition to try to break it, then add more data/variety and test again. Answer the quiz reflections as your Operator Log.

Project Setup

Labels:

Class A: “Close”
Class B: “Far”

Object: Pick ONE object you have (mug, pen, shoe, phone case, spoon).

Safety: No faces. No personal documents. No uniforms. No identifying stuff.

Why this works: it’s not about “recognising a mug”. It’s about learning a visual pattern (size/scale), and then failing when you change conditions.

Step 1: Open the tool

Go to: https://teachablemachine.withgoogle.com/train

Click: Image Project → Standard Image Model

Step 2: Name your classes

Rename:

Class 1 → Close

Class 2 → Far

Step 3: Collect your training data (this is the real skill)

Use webcam capture.

Close (aim for 25 images):

hold the object close to the camera
fill most of the frame

Far (aim for 25 images):

hold the object further back
object is smaller in frame

Important: Keep the background similar at first (same desk/room).
(This makes the “break” and the “scale fix” clearer.)

Step 4: Train

Click Train Model.

Step 5: Test (normal test)

Test it with:

Close object → should say Close
Far object → should say Far

Write down if it makes any mistakes.

Step 6: Break test (Round 2: “New world”)

Now change the conditions (pick ONE):

different background or
different lighting or
different angle

Test again:

If it makes a mistake or gets shaky/confused, great: that’s the point.

Step 7: Scale upgrade (Round 3: add data + variety)

Now you’re going to “level up” the neural network by improving its training data.

Add 10 more images per class (20 total), but this time include variety:

at least 2 different angles
OR a new background
OR different lighting

Then click Train Model again.

Retest (in the new world)

Test the same “new world” conditions again: